Hi, I am an undergraduate student and I am working on a project based on discriminating and characterizing RF partial discharges taken from an electric substation. Its based on four wideband antennas which are placed around the substation which picks up these PD sources in the form of time-domain signals. The signals have been given to me and they look like burst signals (i.e. the signal is silence which then there is burst of energy which is then comes to silence). This is in a space of about 50microseconds. My aim is to apply FFT via matlab on these signals to get the PD's corresponding frequency spectrums. I have to apply windowing techniques and zero-padding in order to achieve the best possible frequency spectrums. The sampling rate is a 4Gsamples/sec; The no. of samples is 2002 where the last 46 points have been padded to zeros, i.e total length is 2048 (Matlab does this automatically); I realise that the best way to get the best possible spectrums is to do the following: 1) multiply data with window coefficients 2) zero pad to next power of 2 3) Perform FFT I have completed these steps but the problem is am not getting the best possible spectrums. I realise the only way I can achieve this is to hone into the actual "burstiness" of the RF signals. For example, since I want to exclude the silence at the beginning and the end of the RF signal I can then window this and FFT it. At this moment the windowing technique is being applied to all the data with the time allocated. The actual signals takes place in a time of 3x10^-7 secs. Would anyone know what would be the best possible way to approach this problem? I have been told that I can zero pad the beginning of the signals, but is this done before or after the burst signal? If this is correct, then how can I do the coding? Thanks SuB-K
FFT on RF Partial Discharge signals
Started by ●March 17, 2008
Reply by ●March 17, 20082008-03-17
Hi, I said in the previous message "I have been told that I can zero pad the beginning of the signals, but is this done before or after the burst signal?" Ignore this, What I meant to say was "I have been told that I can zero pad the beginning of the signals, but is this done before or after the windowing technique?
Reply by ●March 17, 20082008-03-17
SuB-K@hotmail.co.uk wrote:> Hi, I said in the previous message "I have been told that I can zero > pad the beginning of the > signals, but is this done before or after the burst signal?" > Ignore this, What I meant to say was "I have been told that I can zero > pad the beginning of the > signals, but is this done before or after the windowing technique?After. You apply the window to the actual samples, not to any prepended or appended zeros. What are partial discharges? Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������
Reply by ●March 17, 20082008-03-17
On Mar 17, 3:17�pm, Su...@hotmail.co.uk wrote:> Hi, I am an undergraduate student and I am working on a project based > on discriminating and characterizing RF partial discharges taken from > an electric substation. �Its based on four wideband antennas which are > placed around the substation which picks up these PD sources in the > form of time-domain signals. �The signals have been given to me and > they look like burst signals (i.e. the signal is silence which then > there is burst of energy which is then comes to silence). �This is in > a space of about 50microseconds. > My aim is to apply FFT via matlab on these signals to get the PD's > corresponding frequency spectrums. �I have to apply windowing > techniques and zero-padding in order to achieve the best possible > frequency spectrums. > The sampling rate is a 4Gsamples/sec; > The no. of samples is 2002 where the last 46 points have been padded > to zeros, i.e total length is 2048 (Matlab does this automatically); > > I realise that the best way to get the best possible spectrums is to > do the following: > 1) �multiply data with window coefficients > 2) �zero pad to next power of 2 > 3) �Perform FFT > I have completed these steps but the problem is am not getting the > best possible spectrums. �I realise the only way I can achieve this is > to hone into the actual "burstiness" of the RF signals. �For example, > since I want to exclude the silence at the beginning and the end of > the RF signal I can then window this and FFT it. �At this moment the > windowing technique is being applied to all the data with the time > allocated. �The actual signals takes place in a time of 3x10^-7 secs. > Would anyone know what would be the best possible way to approach this > problem? �I have been told that I can zero pad the beginning of the > signals, but is this done before or after the burst signal? �If this > is correct, then how can I do the coding? > > Thanks > SuB-KI would think you DON't want to window these signals since they transient events. If your FFT starts and ends during the "silent" portion of your time data, then there should be no need to window the data. Imagine the limit case where the time data is a single impulse. Mark
Reply by ●March 17, 20082008-03-17
On Mar 17, 9:26 pm, Jerry Avins <j...@ieee.org> wrote:> Su...@hotmail.co.uk wrote: > > Hi, I said in the previous message "I have been told that I can zero > > pad the beginning of the > > signals, but is this done before or after the burst signal?" > > Ignore this, What I meant to say was "I have been told that I can zero > > pad the beginning of the > > signals, but is this done before or after the windowing technique? > > After. You apply the window to the actual samples, not to any prepended > or appended zeros. > > What are partial discharges? > > Jerry > -- > Engineering is the art of making what you want from things you can get. > �����������������������������������������������������������������������Partial Discharges are like "sparks" which occur in insulated high voltage systems (transformers, switchgears, rotating machinery) after years of usages. There are different methods of detecting these "sparks" which in this case its RF acoustic signals. This detection method helps to identify the location/s of the partial discharges therefore maintainance can be applied which can reduce costs and even save lives (remember its high voltages systems).
Reply by ●March 17, 20082008-03-17
On Mar 17, 9:51 pm, Mark <makol...@yahoo.com> wrote:> On Mar 17, 3:17 pm, Su...@hotmail.co.uk wrote: > > > > > Hi, I am an undergraduate student and I am working on a project based > > on discriminating and characterizing RF partial discharges taken from > > an electric substation. Its based on four wideband antennas which are > > placed around the substation which picks up these PD sources in the > > form of time-domain signals. The signals have been given to me and > > they look like burst signals (i.e. the signal is silence which then > > there is burst of energy which is then comes to silence). This is in > > a space of about 50microseconds. > > My aim is to apply FFT via matlab on these signals to get the PD's > > corresponding frequency spectrums. I have to apply windowing > > techniques and zero-padding in order to achieve the best possible > > frequency spectrums. > > The sampling rate is a 4Gsamples/sec; > > The no. of samples is 2002 where the last 46 points have been padded > > to zeros, i.e total length is 2048 (Matlab does this automatically); > > > I realise that the best way to get the best possible spectrums is to > > do the following: > > 1) multiply data with window coefficients > > 2) zero pad to next power of 2 > > 3) Perform FFT > > I have completed these steps but the problem is am not getting the > > best possible spectrums. I realise the only way I can achieve this is > > to hone into the actual "burstiness" of the RF signals. For example, > > since I want to exclude the silence at the beginning and the end of > > the RF signal I can then window this and FFT it. At this moment the > > windowing technique is being applied to all the data with the time > > allocated. The actual signals takes place in a time of 3x10^-7 secs. > > Would anyone know what would be the best possible way to approach this > > problem? I have been told that I can zero pad the beginning of the > > signals, but is this done before or after the burst signal? If this > > is correct, then how can I do the coding? > > > Thanks > > SuB-K > > I would think you DON't want to window these signals since they > transient events. If your FFT starts and ends during the "silent" > portion of your time data, then there should be no need to window the > data. > > Imagine the limit case where the time data is a single impulse. > > MarkIt is part of my thesis to window these signals as a "high-resolution spectrum" can be achieved and where zero-padding will provide a "high- density spectrum" as well. I have a jpeg of the signals but I am unsure how I can put it on this thread for you to see what they look like. What I am trying to achieve is to obtain the best possible frequency spectrums. Applying windows to the whole signal (including the "silent" bits) and then applying an FFT will give me the corresponding spectrum with respect to window applied, be it hamming, bartlett, etc. But if I hone into the actual main signal then I should be able to get an even better spectrum. the problem I am having is how can I do this. The full time-domain signal starts at about -2.5x10^-7 secs and ends at about 2.5x10^-7 secs. The burst signal occurs at 0 secs and eventually dies out at about 2.5x10^-7 secs. In some signals, there is still a little signal left at the end but does not die out completely. I'm assuming that discontinuities can occur here where zero padding will be able to stop this remembering that FFTs repeats itself. I would say that the "silent" bits aren't exactly silent since there are other noise signals from operating machinery within this substation which can cause a bit of distortion but not too much pre 0 seconds (I call this the pre-buffered stage).
Reply by ●March 17, 20082008-03-17
On Mar 17, 9:51 pm, Mark <makol...@yahoo.com> wrote:> On Mar 17, 3:17 pm, Su...@hotmail.co.uk wrote: > > > > > Hi, I am an undergraduate student and I am working on a project based > > on discriminating and characterizing RF partial discharges taken from > > an electric substation. Its based on four wideband antennas which are > > placed around the substation which picks up these PD sources in the > > form of time-domain signals. The signals have been given to me and > > they look like burst signals (i.e. the signal is silence which then > > there is burst of energy which is then comes to silence). This is in > > a space of about 50microseconds. > > My aim is to apply FFT via matlab on these signals to get the PD's > > corresponding frequency spectrums. I have to apply windowing > > techniques and zero-padding in order to achieve the best possible > > frequency spectrums. > > The sampling rate is a 4Gsamples/sec; > > The no. of samples is 2002 where the last 46 points have been padded > > to zeros, i.e total length is 2048 (Matlab does this automatically); > > > I realise that the best way to get the best possible spectrums is to > > do the following: > > 1) multiply data with window coefficients > > 2) zero pad to next power of 2 > > 3) Perform FFT > > I have completed these steps but the problem is am not getting the > > best possible spectrums. I realise the only way I can achieve this is > > to hone into the actual "burstiness" of the RF signals. For example, > > since I want to exclude the silence at the beginning and the end of > > the RF signal I can then window this and FFT it. At this moment the > > windowing technique is being applied to all the data with the time > > allocated. The actual signals takes place in a time of 3x10^-7 secs. > > Would anyone know what would be the best possible way to approach this > > problem? I have been told that I can zero pad the beginning of the > > signals, but is this done before or after the burst signal? If this > > is correct, then how can I do the coding? > > > Thanks > > SuB-K > > I would think you DON't want to window these signals since they > transient events. If your FFT starts and ends during the "silent" > portion of your time data, then there should be no need to window the > data. > > Imagine the limit case where the time data is a single impulse. > > MarkIt is part of my thesis to window these signals as a "high-resolution spectrum" can be achieved and where zero-padding will provide a "high- density spectrum" as well. I have a jpeg of the signals but I am unsure how I can put it on this thread for you to see what they look like. What I am trying to achieve is to obtain the best possible frequency spectrums. Applying windows to the whole signal (including the "silent" bits) and then applying an FFT will give me the corresponding spectrum with respect to window applied, be it hamming, bartlett, etc. But if I hone into the actual main signal then I should be able to get an even better spectrum. the problem I am having is how can I do this. The full time-domain signal starts at about -2.5x10^-7 secs and ends at about 2.5x10^-7 secs. The burst signal occurs at 0 secs and eventually dies out at about 2.5x10^-7 secs. In some signals, there is still a little signal left at the end but does not die out completely. I'm assuming that discontinuities can occur here where zero padding will be able to stop this remembering that FFTs repeats itself. I would say that the "silent" bits aren't exactly silent since there are other noise signals from operating machinery within this substation which can cause a bit of distortion but not too much pre 0 seconds (I call this the pre-buffered stage).
Reply by ●March 17, 20082008-03-17
On Mon, 17 Mar 2008 12:17:47 -0700 (PDT), SuB-K@hotmail.co.uk wrote: (snipped by Lyons)>The sampling rate is a 4Gsamples/sec;(snipped by Lyons)>Thanks >SuB-KHi, is your sampling rate really 4 gigasamples/sec.? [-Rick-]
Reply by ●March 17, 20082008-03-17
On Mar 18, 1:10 am, Rick Lyons <R.Lyons@_BOGUS_ieee.org> wrote:> On Mon, 17 Mar 2008 12:17:47 -0700 (PDT), Su...@hotmail.co.uk wrote: > > (snipped by Lyons) > > >The sampling rate is a 4Gsamples/sec; > > (snipped by Lyons) > > >Thanks > >SuB-K > > Hi, > is your sampling rate really 4 gigasamples/sec.? > > [-Rick-]Yes, if you multiply 2048 by 1/4E^9 you should be able to get 5.12x10^-7 secs. Which should be about the same length of the RF signal within the time-domain.
Reply by ●March 18, 20082008-03-18
SuB-K@hotmail.co.uk wrote:> On Mar 17, 9:26 pm, Jerry Avins <j...@ieee.org> wrote: >> Su...@hotmail.co.uk wrote: >>> Hi, I said in the previous message "I have been told that I can zero >>> pad the beginning of the >>> signals, but is this done before or after the burst signal?" >>> Ignore this, What I meant to say was "I have been told that I can zero >>> pad the beginning of the >>> signals, but is this done before or after the windowing technique? >> After. You apply the window to the actual samples, not to any prepended >> or appended zeros. >> >> What are partial discharges? >> >> Jerry >> -- >> Engineering is the art of making what you want from things you can get. >> ����������������������������������������������������������������������� > > Partial Discharges are like "sparks" which occur in insulated high > voltage systems (transformers, switchgears, rotating machinery) after > years of usages. There are different methods of detecting these > "sparks" which in this case its RF acoustic signals. This detection > method helps to identify the location/s of the partial discharges > therefore maintainance can be applied which can reduce costs and even > save lives (remember its high voltages systems).All right. Sometimes, you can hear the unaided. But why "partial? Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������
